Mythos Business Rationale and The Big Bad Wolf | Sharp Tech with Ben Thompson

0:00 you want.

0:00 You wrote about Anthropic on both Tuesday and Wednesday, and this Mythos model,

0:06 reading about it was pretty unsettling earlier in the week.

0:09 But, what do you think of what's happening here?

0:11 Why is it unsettling?

0:12 We just discussed it on Sharp Tech a week ago.

0:14 Well, we Exactly.

0:16 What we discussed and the threats to security

0:19 that we discussed now appear to be imminent,

0:23 but al- al- be it private for the time being.

0:25 So, I guess we can take solace in that.

0:28 Yo, maybe that's why we're already at stage five.

0:30 Um no, I think a very timely discussion we had last week Mhm.

0:33 about the reality that I mean,

0:37 it's inter- this actually ties into a long-running discussion that we've had,

0:41 particularly this year, about the uniqueness of programming and code

0:47 and its suitability for large language models.

0:50 And the fact that um you know how do you program?

0:54 You put a bunch of words and symbols together in sort of arcane ways

1:00 that it can be difficult for a lot of humans to sort of do,

1:03 but uh human uh computers quite good at it.

1:07 Uh particularly large language models can handle large amounts of language,

1:11 which at the end of the day all of software is just massive amounts of language.

1:15 And again, that language may not be very understandable to you or I,

1:20 but it is very predictable and understandable,

1:23 and this to that end, given what we've talked about,

1:29 this should not be a surprise to to us or to our listeners.

1:33 Um so yeah, here we are.

1:36 Um now how here are we?

1:39 Mhm.

1:41 This is it's it's hard to say.

1:43 Right?

1:44 Because this is Anthropic.

1:46 These are the same people who going back It's funny, people are like,

1:49 "Oh, OpenAI did this, too." No, the Anthropic people did this at OpenAI.

1:55 Where they're like, you know, like,

1:56 "Why is OpenAI not open?" Because GPT-2 posed

2:00 too many dangers to the world, so it's like,

2:03 "Yeah, we're not going to be be be open anymore." And it

2:07 just so happens that not being open is actually good for business.

2:11 And the dangers at that point were what?

2:13 Basically, potentially flooding the world with misinformation?

2:17 Because I mean, that was purely text generation at that point, right?

2:22 very poor text generation, but uh yes, um you know, it's funny,

2:25 we had zero problems with misinformation until GPT-2 came along,

2:28 and then the world has [laughter]

2:31 Still trying to recover from that 2022 release.

2:34 No, I think this is This is like 2019, actually.

2:37 So, this has been a thing for for a long time.

2:40 So and it's also very good for business.

2:43 And and what you think back to the don't So, let me just back up.

2:47 No one get mad at me until I we finish this whole segment, okay?

2:51 Cuz we're going to cover lots of different areas.

2:53 I already see our first email, someone that's very mad at me.

2:56 So, Mr.

2:57 Relax, we're going to [laughter] get there, okay?

3:00 Uh you go back to 2019.

3:02 Uh I think it was 2019 when GPT-2 came out.

3:04 Mhm.

3:05 And there's a this is dangerous.

3:09 And there's also a maybe it's not the best thing in the world

3:12 if we're on the leading edge to give everyone our weights.

3:15 Uh because then they like they can just run the model themselves, right?

3:20 The equivalent here, and by the way,

3:22 I think another area where we were very early,

3:26 what was one of the points that we brought

3:28 up with DeepSeek a year and a half ago?

3:31 Well, DeepSeek looks like it's kind of distilled from leading US models.

3:37 And everyone just sort of takes it as a given or they

3:41 hold it up as an excuse when these labs are complaining about distillation,

3:45 which we've talked about, this idea that you basically query the API

3:48 a gazillion times for all sorts of things,

3:49 and you get your own data from the model to train your own model.

3:53 Like, how do like how do you get these Oh,

3:56 OpenAI is or open source is only 6 months behind.

4:00 Well, cuz it's about 6 months that it

4:01 takes to query these models a gazillion times.

4:04 Successfully distill them.

4:05 Well, can I ask one question on that?

4:07 Because this came up on Sharp China,

4:08 and it's come up a couple different times on Sharp China,

4:11 and I don't have a good answer.

4:13 And as a tech podcaster,

4:14 I feel like I'm failing Bill Bishop in the course of these conversations.

4:18 Is there a way a tough gig for you.

4:21 You you have to be the the dumb normie on this podcast that you have to be the

4:27 [laughter]

4:26 brilliant tech understander on Wearing many hats.

4:30 But, is there a way to reliably prevent distillation in the future?

4:34 Because distilling a model that's as powerful as Mythos

4:37 seems like it could be a problem going forward.

4:40 Yeah, well, I mean, a distilled model is never

4:43 going to be quite as good as the regular one.

4:45 And it's much more jagged, there's much more holes.

4:48 It you know, it's much more much less comprehensive.

4:52 Um so, just in general it's a bit where

4:55 they're always going to be behind to a certain extent.

4:59 But, that doesn't change the fact that if they're more than good enough,

5:02 and these leading edge models are very expensive,

5:04 it's a great alternative if you, you know, want something else.

5:07 So, maybe just to go back to this story there's

5:11 a very good business reason for not making this available,

5:15 just like there's a good reason for not making open weights available.

5:17 This is like it's the same story.

5:20 And and if you think about

5:23 the these companies wanting to have market power/ pricing

5:26 power in the long run well making sure there's not nearly as good models Mhm.

5:34 to the extent you can is a way to do that.

5:37 And the challenge is if you have

5:39 a self-serve walk-up API that anyone can use yeah,

5:42 it's pretty hard it's pretty hard to stop.

5:44 Like, I mean, we've all pirated music.

5:46 It's not like the same story other than

5:49 to say like trying to stop people doing stuff

5:51 on the internet when there's open APIs and things

5:55 that you can access is a tough game.

5:57 Effectively impossible.

5:58 You can make it harder, but not impossible.

6:01 Right.

6:01 And you know, it's one of those things you often find it after it's happened.

6:06 Like, uh you go through your logs and say, "Wow,

6:08 we're getting hit on this endpoint

6:09 from this set of IP addresses a gazillion times,

6:11 which have been routed through a gazillion points, and they're you know,

6:14 it's not like they're coming from like the Forbidden

6:18 City IP range and like accessing the model.

6:21 Like, they're spinning up cloud servers on DigitalOcean or on AWS or whatever,

6:27 and like doing this.

6:28 Probably AWS, probably be too expensive.

6:30 But, like, there's a um it's not Yeah, it's not easy.

6:35 Mhm.

6:35 But, basically, just like it is a rough analogy,

6:38 like policing chips is a lot harder

6:41 than like policing uranium, for example, right?

6:44 Which you can see from satellites,

6:46 and it's much easier to track all over the world.

6:48 like if someone like breaks into like OpenAI and exfiltrates the weights,

6:53 um you know, very clear thievery going on.

6:56 If you're going on and just sort of asking a bunch

6:57 of questions uh at a very high rate of speed,

7:00 which computers are very good at, it's a lot it's a lot tougher to do to stop.

7:05 Yeah.

7:05 Uh so, you have this sort of business issue.

7:08 You also have Anthropic can barely stay online right now.

7:14 [laughter] Mhm.

7:14 you know, the people it it it is this massive upsurge in in revenue, in users.

7:22 They're doing this weird rationing thing like these 5-hour blocks,

7:26 which aren't really 5 hours,

7:27 cuz like the 5 hours is shorter than 5 hours during certain times of day,

7:30 and then it's longer at other times.

7:31 Like, and then people are complaining about, "Oh,

7:34 they're purposely reducing the model quality." There's

7:38 definitely like they're serving distilled models themselves,

7:41 and you can distill much more effectively if you if it's your model,

7:44 and you have like full access to it instead of just using the API.

7:47 And they're quantizing,

7:48 but also they're doing lots of things like trying to leverage cash

7:51 and doing sort of like trying to batch a bunch of stuff together.

7:53 And all these optimizations sort of layer

7:56 on each other to really diminish the experience Mhm.

7:59 To the extent that it's very hard to you so, you have this new model comes out

8:15 that is extremely computationally expensive and intense.

8:19 Uh and you just look at the API pricing, which is like 5x what Opus is.

8:24 And by the way, Opus is significantly more expensive than say GPT-5.5 4,

8:29 which is a even smaller model.

8:31 Um and and so, you it's like if we could limit it not to the hoi polloi,

8:40 but to people who will actually pay us real money,

8:43 also a good sort of business justification, right?

8:45 They'd write you know, and so All of this is making me feel much

8:49 better as we read about a potentially existentially dangerous model here.

8:54 There are lots of rational reasons to approach approach it this way.

8:58 The danger is totally plausible.

9:00 And if And even if the danger in this is why I told everyone to hold off

9:05 the people who want to be mad at me

9:06 even if it's possible they're overstating it right now Yeah.

9:11 it doesn't mean they're overstating the reality

9:14 in 6 months or 9 months or a year.

9:17 The fact of the matter is we are going to have a crisis of thousands,

9:23 not thousands, millions,

9:25 billions of lines of code that have been built by humans from the beginning

9:29 of the computing era till now which

9:32 unquestionably contain tons and tons of bugs,

9:35 cuz that is just the reality of software.

9:38 And theoretically, you could have tons and tons and millions

9:41 of humans go over them and find them all, but that's not practical.

9:46 Mhm.

9:47 But, what are computers really good at?

9:49 Doing boring sort of line by line yeoman's work,

9:53 uh and going over and working through everything.

9:56 And uh the larger these models get and the more capable they

10:00 get and the larger context they have and and the more like,

10:03 yes, this is going to happen.

10:04 So, if it's not happening now and it might be happening now,

10:08 it will be happening in the future.

10:09 So, it's almost pointless to speculate on where Anthropic is with this.

10:15 this.

10:15 Yeah.

10:16 I will I consistently criticize them

10:19 for overstating things where they're at right now.

10:23 And it's a very much a boy but but This is

10:25 why I brought up the boy cried wolf analogy.

10:27 Mhm.

10:28 People talk about the boy crying wolf and they only talk about

10:32 the first 80% of the story where the boy keeps crying wolf.

10:34 [laughter] Yeah.

10:36 At the end of the story, the wolf does come.

10:39 I actually I was not familiar, I mean,

10:41 I'm obviously familiar with the fable there,

10:44 but I didn't know that the wolf does come at the end

10:47 of the boy cried wolf fable until reading Stratechery earlier in the week there.

10:52 So, Wait, what?

10:53 How is that possible?

10:54 it's been what?

10:55 Probably 35 years since I read that story.

10:58 So, over time, I'm familiar with the cliche and not necessarily

11:03 the original text undergirding the cliche that we all know and love.

11:08 all of those fables, like the real versions, are all very like dark.

11:12 Dark?

11:14 [laughter] Well, Germanic Arguably something we've forgotten, right?

11:16 The point of them was to instill

11:19 healthy fear and instincts into children, right?

11:23 Like Yeah.

11:25 [laughter] You know, there's there's been a real movement to soften

11:28 all these things and make them more complex and

11:31 Oh, believe me, I'm reading children's book children's

11:33 books every single night and nobody ever dies, nothing bad ever happens.

11:38 Whereas my wife grew up with her mom

11:41 reading her like German fables where it's really grizzly.

11:46 [laughter] Um so, perhaps we're at a better place on that front or perhaps not.

11:51 Perhaps children needed those lessons from the Germans way back when.

11:55 There's that's that might be the case.

11:56 Well,